Semantically Contrastive Learning for Low-light Image Enhancement

Introduction

• Underexposed images remain a significant source of camera-imaging errors and can hinder semantic segmentation and object detection in vision-based driving assistance.Source: Introduction · p. 1
• Unpaired over/underexposed images are readily available, while high-quality normal-light positives can be difficult to acquire and may require complicated registration and calibration.The paper motivates using these images as negative samples to provide supervisory information for visual mapping.Source: Introduction · p. 1
• Existing methods overlook scene semantics and within-category brightness consistency, although these can guide distinct foreground/background enhancement and help avoid uneven exposure.The paper notes that enhancement strategies should differ between background and foreground areas.Source: Introduction · p. 1
• SCL-LLE casts enhancement as joint learning with contrastive learning, semantic brightness consistency, and feature preservation constraints for exposure, texture, and color consistency.It uses unpaired normal-light positives and over/underexposed negatives alongside semantic information.Source: Introduction · p. 2
• SCL-LLE removes pixel-corresponding paired training data and is evaluated across six independent datasets using visual, image-quality, and human assessments.The paper reports that the results consistently endorse the approach's superiority.Source: Introduction · p. 2

Low-light Image Enhancement

• Handcrafted priors such as histogram equalization and Retinex are not self-adaptive enough to recover image details and color, with artifacts including uneven exposure and halos.Source: Low-light Image Enhancement · p. 2
• Many deep LLE methods use strictly pixel-corresponding image pairs, while Jiang et al. (2021) and Zero-DCE avoid paired low-light references in different ways.The paper distinguishes its method by learning brightness from unpaired negative samples.Source: Low-light Image Enhancement · p. 2
• Fan et al. (2020) use semantic information to guide Retinex reflection reconstruction for noise removal, whereas this work applies semantics to brightness reconstruction and semantic-element dependencies.Source: Low-light Image Enhancement · p. 2

Contrastive Learning

• Contrastive learning pulls representations toward positives and away from negatives; SCL-LLE uses a couple of negatives per positive and random mapping to reduce overfitting risk.The paper presents this approach as a way to learn from available open data without large negative-sample collections.Source: Contrastive Learning · p. 2

Problem Formulation and Overall Architecture

• SCL-LLE maps low-light inputs to enhanced images using exposure-sample and semantic priors within a framework of enhancement, segmentation, and feature-extraction networks.The three losses correspond to contrastive learning, semantic brightness consistency, and feature preservation.Source: Problem Formulation and Overall… · pp. 2–3, Figure 2 · p. 3

Contrastive Brightness Restoration

• Normal-light images are positive samples and over- or underexposed images are negatives, unpaired both with the input and with each other.Training randomly selects positive and negative samples; the contrastive objective brings enhanced-image features closer to positive features than negative features.Source: Contrastive Brightness Restoration · pp. 3–4

Semantic Brightness Consistency Constraint

• Semantic brightness consistency regularizes brightness within each semantic category to help avoid local over- or underexposure.The constraint uses average brightness for pixels in each ground-truth category and the semantic prediction.Source: Semantic Brightness Consistency… · p. 4

Feature Preservation

• Feature preservation combines feature retention with color constancy to preserve input features and constrain color deviations and illumination changes.Feature retention compares input and enhanced-image feature maps; the color term incorporates channel-ratio constraints and illumination smoothness.Source: Feature Preservation · p. 4

Efficient Training Details

• Training fixes the segmentation and feature-extraction network weights, updating only the enhancement network; images are resized to 384×384 and trained for 50 epochs with batch size 2.The reported setup uses Adam with a fixed learning rate of 1e−4 on an NVIDIA 1080Ti GPU.Source: Efficient Training Details · pp. 4–5

Cross-dataset Peer Comparison

• In visual comparisons, SCL-LLE recovers more detail and contrast in foreground and background, while competing methods show exposure, color, or detail defects.The evaluation uses six independent low-light datasets and compares SCL-LLE with six representative methods.Source: Visual quality comparison · p. 5, Cross-dataset Peer Comparison · p. 5
• SCL-LLE has the best overall averaged NIQE and UNIQUE results, the best PSNR and SSIM on SICE Part2, and the best human-survey imaging quality.The reported NIQE and UNIQUE comparisons cover five datasets; the survey ranks enhanced images from those datasets.Source: No-referenced image quality assessment · p. 6, Table 1 · p. 6, Full-referenced image quality assessment · p. 6

Ablation Study

• Removing semantic brightness consistency or feature retention yields lower contrast than the final model, while the full losses produce fine details and smoother global intensity distributions.The ablation reports undesirable unilateral over- or under-saturation in the histograms for reduced-loss settings.Source: Contribution of each loss · p. 6, Table 3 · p. 6, Figure 5 · p. 7
• Removing overexposed or underexposed negative samples shifts white balance in opposite directions, supporting the reported necessity of both types during training.The tested settings include positive samples alone, either negative type alone, and both types together.Source: Impact of negative samples · pp. 6–7, Figure 5 · p. 7

Semantic Segmentation with LLE

• On Cityscapes, enhancement methods do not surpass the original input at γ=1, but at larger γ they improve segmentation mIoU, with SCL-LLE tending to perform best in darker scenes.The test uses gamma-transformed images because current low-light datasets lack semantic annotations.Source: Semantic Segmentation with LLE · p. 7, Table 4 · p. 7

• Experiments show clear improvements over existing state-of-the-art LLE models on six cross-domain datasets.Source: Conclusion · p. 7
• SCL-LLE combines contrastive learning, semantic brightness consistency, and feature preservation to ensure color, texture, and exposure consistency.Source: Conclusion · p. 7
• The method uses unpaired positive and negative samples and semantic information to produce visually pleasing images while preserving similarity between input and output.Source: Conclusion · p. 7
• Experiments indicate that SCL-LLE has potential to improve downstream semantic segmentation performance.Source: Conclusion · p. 7